Skip to content
SDGs Center UDINUS

Evidence record · UDINUS · SDGs

IMPACT OF HISTORICAL DATA LENGTH ON LSTM AND SINGLE BLOCK ENCODER-DECODER TRANSFORMER FOR MULTI-STEP TRAFFIC FLOW FORECASTING

Rektorat UDINUS 2026 Published 2026-10-30

Evidence record · UDINUS · SDGs

Unit
Rektorat UDINUS
Period
2026
Goals
Record
42d20711-d749-46b2-aa3c-fa66f8cf5a3d
Published
2026-10-30

Universitas Dian Nuswantoro conducted an activity examining the impact of historical data length on multi-step traffic flow forecasting. The activity used two deep learning models, namely LSTM and Single Block Encoder-Decoder Transformer, to explore how much historical data is needed to achieve accurate predictions. The forecasting application involved hourly traffic volume and a one-day-ahead forecast. The scope included four historical input lengths: 1, 2, 3, and 4 weeks. The dataset was the I5-North Freeway District 10, California dataset. Activity participants listed for the work were Omar, Mas; Yakub, Fitri; Azizan, Azizul; Mamat, Normaisharah; Dhamanti, Inge; Govindan, Niranjana; Chinnathambi, Dhaya; Muljono, Muljono; and Monterio, Cancio.

For both models, a 2-week historical input length yielded the best overall accuracy under the Combined Accuracy metric. One week of data worked but provided insufficient context. Using 3 and 4 weeks appeared to provide unnecessary information to both models, which harmed accuracy because of overfitting and additional noise. The activity therefore indicates that adding more historical input data beyond a certain length does not necessarily improve predictions and can make predictions worse. The work addresses traffic volume forecasting and effective traffic management by examining historical inputs, deep learning models, and prediction accuracy. Further information is not provided about the research procedures, study outcomes beyond the reported accuracy comparison, or institutional implementation of the findings.

Reference https://www.scopus.com/pages/publications/105031821849?origin=resultslist